Deep learning interprets failure process of coal reservoir during CO2-desorption by 3D reconstruction techniques

This work aims to investigate the failure and mechanical behaviors of coal rocks during CO2-desorption. The unconfined and confined compression experiments with certain depletion pressure-stress path and desorption of CO2 are conducted, in which the cracking process is captured by X-ray CT imaging. A deep learning-based coal fine-crack segmentation (CFCS) network with 3D reconstruction technique is proposed to study the cracking behaviors. In addition, the CO2 evolution effects on the failure process and mechanical responses are also analyzed. Results show that the proposed CFCS can effectively segment the crack-CO2-soid structure in coal with the errors all less than 10% at the range of 3.5620%–9.9967%. The failure process of coal samples can be divided into the extremely fast CO2-desorption + microcracks formation stage, fast CO2-desorption + stable microcracks propagation stage and slow stable CO2-desorption + microcracks coalescence to form macrocracks stage. The shear failures of coal samples are caused by the reduction of lateral effective stress due to CO2-desorption and drainage, which increases the anisotropy of coal samples with microcracking behaviors. The proposed method provides better understanding of failure mechanism and mechanical responses of coal reservoirs, which are helpful for the CO2 storage-based resource explorations.